Abstract: Aiming at the problems of laser vision robots during the welding process,the laser stripe segmentation accuracy of the weld seam was reduced due to noise interference and the inaccurate welding,an improved U-Net weld seam feature recognition method was proposed. The improved U-Net algorithm incorporates Mobile-Net as the underlying network,thereby augmenting the network′s capability for feature recognition and reducing the number of model parameters. Efficient channel attention network was added between encoding and decoding to achieve weighted fusion of features.The model adopted a mixed loss function to balance the proportion of laser stripes in the image. The network model was implemented on the experimental platform for seam tracking in welding robots. Experimental results showed that the Mean Intersection over Union (MIoU) of the improved U-Net algorithm was 89.83 %,the Pixel Accuracy (PA) was 99.54 %,the Mean Pixel Accuracy (MPA) was 97.28 %,and the image processing time was 0.209 s.Compared with other algorithms,it has better segmentation accuracy and faster processing speed.It can be more effectively applied to robot welding scenarios with interference.

Key words: image segmentation, robot welding, U-Net algorithm, ECA attention mechanism, weld feature recognition